Jev_Qwen3.8-27B-r2-LoRA

Round-2 calibrated judge adapter — the 27B trained on data it chose for itself.

LoRA adapter (r=64, alpha=128) for huihui-ai/Huihui-Qwen3.8-27B-abliterated. This is the adapter ONLY — it must be applied to that exact base (see serve command below; do not peft-merge, see warning).

What it does

Teaches the base to be a calibrated judge: brief calls with honest, stated confidence — and genuine "roughly a toss-up" behavior on 50/50 questions instead of false certainty.

Measured vs round-1 tune on the same eval battery: toss-up overconfidence 15/20 → 0-1/20 (round-1 answered genuine 50/50s with ≥90% certainty; round-2 answers "roughly a toss-up, 40-60%", committing only where evidence leans). Everyday decision style preserved from base.

Full lineage: SargeDev/Jev_Qwen3.8-27B (round-1 merge, published) — trained on SargeDev/jev-distill-corpus-v3.

⚠️ Do not peft-merge this adapter

merge_and_unload (bf16 or fp32 accumulation) produces a model that outputs uniform 0.5/0.5 for everything — verified broken. The adapter works correctly only via runtimes that apply LoRA deltas in their own kernels: vLLM is verified working. Serving the merged result is not supported.

Serving (verified on DGX Spark / GB10, vLLM)

BASE=/path/to/huihui-ai/Huihui-Qwen3.8-27B-abliterated
vllm serve "$BASE" \
  --enable-lora --max-lora-rank 64 \
  --lora-modules tuned=/path/to/this/adapter \
  --served-model-name jev-r2 \
  --trust-remote-code --reasoning-parser qwen3 \
  --enable-chunked-prefill --enable-prefix-caching \
  --max-model-len 8192 --gpu-memory-utilization 0.80

Then request model: "tuned". Pass "chat_template_kwargs": {"enable_thinking": false} for tuned behavior (trained with thinking off).

Training summary

51,663 rows — the model itself selected its training data: sampled its own predictions over 4,779 held-out corpus rows and kept the rows where its calibration was broken (confident-conflicts 1,614 + toss-up band 1,044 + uncertain 169; wild-disagreements dropped)
QLoRA: r64/α128 on NF4 (double-quant) base, completion-only loss, paged_adamw_8bit
checkpoint-300 of 900 steps (eval loss 0.432 at 300 vs 0.480 at 900 — less overfit; A/B vs final showed identical behavior)
Trained on NVIDIA DGX Spark (GB10), 25h

Apache-2.0.

Training dataset

Trained on SargeDev/jev-distill-corpus-v3 — 740,957 rows of typed calibrated decisions (noul/choice/score) distilled from Jev 1.13 (System One).

SargeDev/Jev_Qwen3.8-27B-r2-LoRA